Particle Swarm Optimization of a Liquid State Machine for Epilepsy and Seizure EEG Signal Classification

Abstract

Electroencephalogram (EEG) data classification is still a complex and time-consuming task so far. In this paper, a liquid state machine (LSM), a bio-inspired computing model, was investigated for classification of the epileptic seizure EEG data. To enhance the classification performance of the random connected LSM, the particle swarm optimization (PSO) algorithm was used to optimize its topology and nonlinear dynamics through the search of the liquid hyperparameters such as the scaling of synaptic strength, connection probability and time constant of membrane potential. The effect of the inertia weight in PSO on the performance of searched LSM was studied. The best accuracy of 95% for EEG classification was achieved by an optimized LSM with 160 neurons combined with a Softmax classifier via 10-fold cross-validation. In addition, the margin of the searched parameters for hardware implementation of LSM was given.